1. Field of the Invention
The present invention relates generally to object inspection apparatus and more particularly to an automatic photomask inspection system and apparatus capable of locating extremely small defects in the transparencies of a photomask used to manufacture semiconductor devices.
2. Discussion of the Prior Art
One of the major sources of yield loss in the manufacture of large scale integrated circuits (LSI) is random defect in the photomasks as chip sizes and geometry densities increase more attention must be given to mask quality in order to reduce defects to a level commensurate with acceptable yields. Accordingly, various methods of detection of photomask defects have been utilized in the past including the comparison of the subject mask to a known reference, the use of an image enhancement technique, and the use of a technique wherein adjacent die are compared to each other.
In order to determine how mask defects may be reduced it is first necessary to examine defect sources. These may be conveniently divided into three categories: (1) Defects present in the manufacture of the mask, (2) defects added to the mask during handling and storage or transportation and (3) defects added to the mask during mask-to-wafer alignment and exposure. To date a great deal of effort and capital has been expended on category (3) defects, the industry has abandoned vacuum contact printing to reduce add on defects introduced during alignment. Handling and storage defects have been reduced by better mask carriers. Additionally, automated mask cleaners are available to remove some of the contamination added to the mask during its movements about the plant or use in an aligner.
Presently, with the industry-wide move to projection printing, increased attention is being focused on category (1) defects, those defects which are in the mask as originally manufactured. These defects have been reduced in the mask shop by placing tighter controls on the mask blanks, reducing the number of steps in the process, i.e., eliminating submaster steps, and inspecting and repairing master and submaster masks.
Mask inspection is a difficult task. For example, inspecting 5".times.5" array for 50 microinch defects means testing 10.sup.10 discrete location for conformity with the required image. A human operator takes many hours to complete such a task and is subject to fatigue. Additionally difficulty arises since defect criteria are subjective and small defects may go unnoticed. When the inspector is under pressure from volume and yield requirements, it has been shown that there may be considerable divergence between the results he obtains and an unbiased test on the same mask. Even in unbiased tests of this type many of the smaller defects may be missed. The common practice of statistical sampling to determine defect density further decreases the confident level in the defect density measurement obtained by a human operator. If mask repair of mask matching is to be done, 100% testing is, of course, necessary as the exact location of each defect must be found.
In utilizing the first above-mentioned technique, an optical comparator is used to compare a test mask to a reference mask. The two masks are mounted on a common stage and are illuminated from below. The transmitted light is filtered to give complementary colors and the two images are superimposed so that common areas are black or white while differences appear as colored areas. The two masks have to be exactly aligned to each other and individually focused over the entire area to be examined. This may take considerable time and skill. Small edge misalignments cause edge coloring but a human operator can usually detect the differences between defects and edge registration Misregistration becomes a major problem if one attempts to automate such a system. Furthermore, since the two images are widely separated in space, the mechanical requirements on the stage are stringent which inevitably results in high costs. Due to the cost alignment time and tediousness of the process, systems of this type are normally utilized as overlay comparator to check for level-to-level mask fit and line width variations rather than random defects. Extending this method to general defect detection involves processing and storage of an enormous amount of data and processing and computing speed must also be very high for reasonable inspection times to be achieved.
The image enhancement technique makes use of the fact that most valid semiconductor geometries are orthogonal whereas defective geometries are randomly oriented. To take advantage of this difference in properties between the semiconductor geometries and the defect, optical equipment is utilized which discriminates against horizontal and vertical edges by reducing their intensity relative to edges at other angles. This result can be achieved by taking advantage of the fact that when an object is illuminated by coherent light, due to diffraction the Forier spectrum of the object is formed at the back focal plane of the imaging lens. By placing horizontal and vertical stops (spacial filters) at the back focal plane of the lens, most of the energy from the periodic horizontal and vertical structure can be blocked. This decreases the energy contribution from ths structure to the final image. Nonhorizontal and vertical structures have most of the energy in other planes and are not effected as much by these spacial filters. A variation of this concept is to place spacial filters in a noncoherent illuminator of a standard microscope. The image of the spacial filters in the illuminator lies close to the back focal plane of the imaging optics, then the system is very similar to that obtained by placing spacial filters in the imaging system. This method yields somewhat inferior results compared to the coherent illumination system but has the advantage of lower costs. These methods can be extended with some success to cover edges occurring at other common angles, i.e., 45.degree. and 60.degree., and considerable visual enhancement of the defects can thus be obtained. Unfortunately, the corners of geometries contain components at every angle within their quadrant and cannot therefore be completely removed by these techniques; thus, while these methods do enable the defects to be more easily detected by the eye, they do not lend themselves easily to automatic detection.
Since masks have the unique property of repetition, all die should have the same pattern. If a difference occurs in the comparison of one die to another, comparison of either die to a third will determine which die is defective. This, of course, assumes that the defect is not repeated in the same place in the several die. There is negligible probability of this occurring unless the defect was present in the original reticle. For example, if there are 10 defects per square inch and the average defect size measures 0.001".times.0.001" the probability of two random defects occurring in the same spot and thus being missed is one tenth in 100,000. At this defect size and density, only one defect would be missed every 600 masks.
The comparison between adjacent die method of inspection is most readily adaptable to fast automatic inspection. There is only one mask to set up and focus, no reference is required and because the patterns are on the same mask, the distance between them is small and fixed. This eases the mechanical tolerances of the table required to scan the mask. For production use, automatic focus is a desirable feature, particularly for inspection of large masks where the bow of the glass due to the mask weight must be considerable compared to the length of focus of the optical system.
A difficult requirement to meet even when using the comparison between adjacent die method of detection defects is optical and mechanical alignment the predision to which the two sets of optics must be placed over the circuits and the precision with which the stage must move determines the smallest defect this typed of system can detect. This can readily be seen since a misregistration of the images from the two geometries under test results in the system identifying each misregistered portion as a defect; therefore, if the mechanics of the system allow misregistration of one micron, then the system logic must reject defects smaller than one micron to avoid confusion between actual defects and misregistered valid geometry.